Trustworthy Abnormality Detection from Welding Images Through Class-Conditional Conformal Learning and Bayesian Cost Minimization

Abstract

Industrial fault/abnormality detection is often criticized for lacking explainability and robustness. Furthermore, practical industrial datasets are frequently highly imbalanced and operate under extreme risk asymmetry, i.e., false negatives carry penalties orders of magnitude higher than false alarms, which poses significant challenges to reliable detection. In this paper, we develop a trustworthy AI framework to improve confidence in welding defect detection. The proposed framework integrates two primary techniques: class-conditional conformal learning and Bayesian cost minimization. First, the conformal learning model quantifies the trustworthiness of model predictions. Instead of forcing a binary classification, the model outputs an "uncertain" state when confidence is low, facilitating informed human intervention. Second, a Bayesian cost minimization algorithm is used to avoid over-conservative predictions that yield too many "uncertain" predictions. Results on a real-world welding quality inspection dataset show that the developed method adapts robustly to dynamic intervention costs and mitigates worst-case cost spikes. The framework is deployment-oriented: it is not uniformly optimal in every setting, but it consistently avoids catastrophic failures and maintains a favorable cost–accuracy–intervention trade-off across heterogeneous base-model qualities.